CausalVAD: De-confounding End-to-End Autonomous Driving via Causal Intervention

2026-03-19

AI Summary

End-to-end autonomous driving models are prone to dataset biases, often learning spurious statistical correlations that compromise safety and reliability in complex scenarios. To address this, this work proposes CausalVAD, a novel framework that integrates lightweight backdoor adjustment theory into end-to-end driving systems for the first time, introducing a plug-and-play Sparse Causal Intervention Scheme (SCIS). SCIS leverages a driving context prototype dictionary to sparsely intervene on query vectors, effectively removing confounding biases from learned representations and enhancing the causal reliability of motion planning. Evaluated on benchmarks such as nuScenes, CausalVAD achieves state-of-the-art performance in both planning accuracy and safety, while demonstrating exceptional robustness under data bias and sensor noise.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Planning-oriented end-to-end driving models show great promise, yet they fundamentally learn statistical correlations instead of true causal relationships. This vulnerability leads to causal confusion, where models exploit dataset biases as shortcuts, critically harming their reliability and safety in complex scenarios. To address this, we introduce CausalVAD, a de-confounding training framework that leverages causal intervention. At its core, we design the sparse causal intervention scheme (SCIS), a lightweight, plug-and-play module to instantiate the backdoor adjustment theory in neural networks. SCIS constructs a dictionary of prototypes representing latent driving contexts. It then uses this dictionary to intervene on the model's sparse vectorized queries. This step actively eliminates spurious associations induced by confounders, thereby eliminating spurious factors from the representations for downstream tasks. Extensive experiments on benchmarks like nuScenes show CausalVAD achieves state-of-the-art planning accuracy and safety. Furthermore, our method demonstrates superior robustness against both data bias and noisy scenarios configured to induce causal confusion.

Problem

• causal confusion
• end-to-end autonomous driving
• dataset bias
• spurious correlations
• confounders

Innovation

• causal intervention
• de-confounding
• end-to-end autonomous driving
• backdoor adjustment
• sparse causal intervention scheme